libvpx: Pull from upstream

source/libvpx/vp9/common/vp9_convolve.c

Index: source/libvpx/vp9/common/vp9_convolve.c
===================================================================
--- source/libvpx/vp9/common/vp9_convolve.c	(revision 240950)
+++ source/libvpx/vp9/common/vp9_convolve.c	(working copy)
@@ -18,40 +18,21 @@
 #include "vpx/vpx_integer.h"
 #include "vpx_ports/mem.h"
 
-static void convolve_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
-                             uint8_t *dst, ptrdiff_t dst_stride,
-                             const int16_t *filter_x0, int x_step_q4,
-                             const int16_t *filter_y, int y_step_q4,
-                             int w, int h, int taps) {
-  int x, y, k;
-
-  /* NOTE: This assumes that the filter table is 256-byte aligned. */
-  /* TODO(agrange) Modify to make independent of table alignment. */
-  const int16_t *const filter_x_base =
-      (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
-
-  /* Adjust base pointer address for this source line */
-  src -= taps / 2 - 1;
-
+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const subpel_kernel *x_filters,
+                           int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
   for (y = 0; y < h; ++y) {
-    /* Initial phase offset */
-    int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
-
+    int x_q4 = x0_q4;
     for (x = 0; x < w; ++x) {
-      /* Per-pixel src offset */
-      const int src_x = x_q4 >> SUBPEL_BITS;
-      int sum = 0;
-
-      /* Pointer to filter to use */
-      const int16_t *const filter_x = filter_x_base +
-          (x_q4 & SUBPEL_MASK) * taps;
-
-      for (k = 0; k < taps; ++k)
-        sum += src[src_x + k] * filter_x[k];
-
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
       dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-
-      /* Move to the next source pixel */
       x_q4 += x_step_q4;
     }
     src += src_stride;
@@ -59,41 +40,22 @@
   }
 }
 
-static void convolve_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
-                                 uint8_t *dst, ptrdiff_t dst_stride,
-                                 const int16_t *filter_x0, int x_step_q4,
-                                 const int16_t *filter_y, int y_step_q4,
-                                 int w, int h, int taps) {
-  int x, y, k;
-
-  /* NOTE: This assumes that the filter table is 256-byte aligned. */
-  /* TODO(agrange) Modify to make independent of table alignment. */
-  const int16_t *const filter_x_base =
-      (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
-
-  /* Adjust base pointer address for this source line */
-  src -= taps / 2 - 1;
-
+static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const subpel_kernel *x_filters,
+                               int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
   for (y = 0; y < h; ++y) {
-    /* Initial phase offset */
-    int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
-
+    int x_q4 = x0_q4;
     for (x = 0; x < w; ++x) {
-      /* Per-pixel src offset */
-      const int src_x = x_q4 >> SUBPEL_BITS;
-      int sum = 0;
-
-      /* Pointer to filter to use */
-      const int16_t *const filter_x = filter_x_base +
-          (x_q4 & SUBPEL_MASK) * taps;
-
-      for (k = 0; k < taps; ++k)
-        sum += src[src_x + k] * filter_x[k];
-
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
       dst[x] = ROUND_POWER_OF_TWO(dst[x] +
-                   clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
-
-      /* Move to the next source pixel */
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
       x_q4 += x_step_q4;
     }
     src += src_stride;
@@ -101,41 +63,22 @@
   }
 }
 
-static void convolve_vert_c(const uint8_t *src, ptrdiff_t src_stride,
-                            uint8_t *dst, ptrdiff_t dst_stride,
-                            const int16_t *filter_x, int x_step_q4,
-                            const int16_t *filter_y0, int y_step_q4,
-                            int w, int h, int taps) {
-  int x, y, k;
+static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const subpel_kernel *y_filters,
+                          int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
 
-  /* NOTE: This assumes that the filter table is 256-byte aligned. */
-  /* TODO(agrange) Modify to make independent of table alignment. */
-  const int16_t *const filter_y_base =
-      (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
-
-  /* Adjust base pointer address for this source column */
-  src -= src_stride * (taps / 2 - 1);
-
   for (x = 0; x < w; ++x) {
-    /* Initial phase offset */
-    int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
-
+    int y_q4 = y0_q4;
     for (y = 0; y < h; ++y) {
-      /* Per-pixel src offset */
-      const int src_y = y_q4 >> SUBPEL_BITS;
-      int sum = 0;
-
-      /* Pointer to filter to use */
-      const int16_t *const filter_y = filter_y_base +
-          (y_q4 & SUBPEL_MASK) * taps;
-
-      for (k = 0; k < taps; ++k)
-        sum += src[(src_y + k) * src_stride] * filter_y[k];
-
-      dst[y * dst_stride] =
-          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-
-      /* Move to the next source pixel */
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
       y_q4 += y_step_q4;
     }
     ++src;
@@ -143,41 +86,23 @@
   }
 }
 
-static void convolve_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
-                                uint8_t *dst, ptrdiff_t dst_stride,
-                                const int16_t *filter_x, int x_step_q4,
-                                const int16_t *filter_y0, int y_step_q4,
-                                int w, int h, int taps) {
-  int x, y, k;
+static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const subpel_kernel *y_filters,
+                              int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
 
-  /* NOTE: This assumes that the filter table is 256-byte aligned. */
-  /* TODO(agrange) Modify to make independent of table alignment. */
-  const int16_t *const filter_y_base =
-      (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
-
-  /* Adjust base pointer address for this source column */
-  src -= src_stride * (taps / 2 - 1);
-
   for (x = 0; x < w; ++x) {
-    /* Initial phase offset */
-    int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
-
+    int y_q4 = y0_q4;
     for (y = 0; y < h; ++y) {
-      /* Per-pixel src offset */
-      const int src_y = y_q4 >> SUBPEL_BITS;
-      int sum = 0;
-
-      /* Pointer to filter to use */
-      const int16_t *const filter_y = filter_y_base +
-          (y_q4 & SUBPEL_MASK) * taps;
-
-      for (k = 0; k < taps; ++k)
-        sum += src[(src_y + k) * src_stride] * filter_y[k];
-
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
       dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
-           clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
-
-      /* Move to the next source pixel */
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
       y_q4 += y_step_q4;
     }
     ++src;
@@ -185,42 +110,54 @@
   }
 }
 
-static void convolve_c(const uint8_t *src, ptrdiff_t src_stride,
-                       uint8_t *dst, ptrdiff_t dst_stride,
-                       const int16_t *filter_x, int x_step_q4,
-                       const int16_t *filter_y, int y_step_q4,
-                       int w, int h, int taps) {
-  /* Fixed size intermediate buffer places limits on parameters.
-   * Maximum intermediate_height is 324, for y_step_q4 == 80,
-   * h == 64, taps == 8.
-   * y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
-   */
+static void convolve(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const subpel_kernel *const x_filters,
+                     int x0_q4, int x_step_q4,
+                     const subpel_kernel *const y_filters,
+                     int y0_q4, int y_step_q4,
+                     int w, int h) {
+  // Fixed size intermediate buffer places limits on parameters.
+  // Maximum intermediate_height is 324, for y_step_q4 == 80,
+  // h == 64, taps == 8.
+  // y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
   uint8_t temp[64 * 324];
-  int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + taps;
+  int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
 
   assert(w <= 64);
   assert(h <= 64);
-  assert(taps <= 8);
   assert(y_step_q4 <= 80);
   assert(x_step_q4 <= 80);
 
   if (intermediate_height < h)
     intermediate_height = h;
 
-  convolve_horiz_c(src - src_stride * (taps / 2 - 1), src_stride, temp, 64,
-                   filter_x, x_step_q4, filter_y, y_step_q4, w,
-                   intermediate_height, taps);
-  convolve_vert_c(temp + 64 * (taps / 2 - 1), 64, dst, dst_stride, filter_x,
-                  x_step_q4, filter_y, y_step_q4, w, h, taps);
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
+                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
+  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
+                y_filters, y0_q4, y_step_q4, w, h);
 }
 
+static const subpel_kernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const subpel_kernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const subpel_kernel *base) {
+  return (const subpel_kernel *)(intptr_t)f - base;
+}
+
 void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
                            const int16_t *filter_y, int y_step_q4,
                            int w, int h) {
-  convolve_horiz_c(src, src_stride, dst, dst_stride,
-                   filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
+  const subpel_kernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                 x0_q4, x_step_q4, w, h);
 }
 
 void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
@@ -228,8 +165,11 @@
                                const int16_t *filter_x, int x_step_q4,
                                const int16_t *filter_y, int y_step_q4,
                                int w, int h) {
-  convolve_avg_horiz_c(src, src_stride, dst, dst_stride,
-                       filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
+  const subpel_kernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                     x0_q4, x_step_q4, w, h);
 }
 
 void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
@@ -237,8 +177,10 @@
                           const int16_t *filter_x, int x_step_q4,
                           const int16_t *filter_y, int y_step_q4,
                           int w, int h) {
-  convolve_vert_c(src, src_stride, dst, dst_stride,
-                  filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
+  const subpel_kernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                y0_q4, y_step_q4, w, h);
 }
 
 void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
@@ -246,8 +188,10 @@
                               const int16_t *filter_x, int x_step_q4,
                               const int16_t *filter_y, int y_step_q4,
                               int w, int h) {
-  convolve_avg_vert_c(src, src_stride, dst, dst_stride,
-                      filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
+  const subpel_kernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                    y0_q4, y_step_q4, w, h);
 }
 
 void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
@@ -255,8 +199,15 @@
                      const int16_t *filter_x, int x_step_q4,
                      const int16_t *filter_y, int y_step_q4,
                      int w, int h) {
-  convolve_c(src, src_stride, dst, dst_stride,
-             filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
+  const subpel_kernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const subpel_kernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride,
+           filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
 }
 
 void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
@@ -269,9 +220,9 @@
   assert(w <= 64);
   assert(h <= 64);
 
-  vp9_convolve8(src, src_stride, temp, 64,
-               filter_x, x_step_q4, filter_y, y_step_q4, w, h);
-  vp9_convolve_avg(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
+  vp9_convolve8_c(src, src_stride, temp, 64,
+                  filter_x, x_step_q4, filter_y, y_step_q4, w, h);
+  vp9_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
 }
 
 void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,